#8763: forM_ [1..N] does not get fused (allocates 50% more) -------------------------------------+------------------------------------- Reporter: nh2 | Owner: (none) Type: bug | Status: new Priority: normal | Milestone: 8.8.1 Component: Compiler | Version: 7.6.3 Resolution: | Keywords: Operating System: Unknown/Multiple | Architecture: Type of failure: Runtime | Unknown/Multiple performance bug | Test Case: Blocked By: | Blocking: Related Tickets: #7206 | Differential Rev(s): Wiki Page: | -------------------------------------+------------------------------------- Comment (by simonpj):
I didn't know that INLINE works 'both ways', e.g. that we can prevent inlining the body into c this way.
Yes, it's super-important that INLINE works like that. Consider {{{ f x = <very big> g y = Just (f y) {-# INLINE g #-} }}} where that is the only occurrence of `f`, but there are zillions of occurrences of `g`. You would be jolly annoyed if GHC inlined `<very big>` into `g`, and then inlined the new `g` at each of its zillions of call sites!! No: INLINE says "when you see a saturated application of this function, inline ''the RHSI wrote'' at the call site". It might be useful to beef up the documentation of INLINE to explain this, if you felt able to do so. I agree that there may be other places such an INLINE would be desirable. A good start would be a careful Note on `mapM_` explaining from first principles why the local let and INLINE is important. Then others can refer to that Note. -- Ticket URL: http://ghc.haskell.org/trac/ghc/ticket/8763#comment:78 GHC http://www.haskell.org/ghc/ The Glasgow Haskell Compiler